Grandet finds the cheapest API price for any model from your agent
Grandet is a free, anonymous MCP server and agent skill that reads prices from more than 1,800 model APIs across official vendors, inference providers and relay sites, and connects the cheapest one to a coding agent.
Image: GitHub
Why it mattersAn agent that can ask for the lowest listed price on the exact Claude or GPT model it is already using is a cheaper monthly bill in exchange for one extra tool call per week.
Anyone paying a monthly bill for a coding agent's API spends time working through the pricing tables of competing providers. Grandet is a free MCP server and agent skill that reads the listed prices for more than 1,800 model APIs, from official vendors through inference providers to relay sites, and tells a coding agent the cheapest few on the exact model it is already using. The project is at github.com/hanzohanzhe/Grandet, the website is at grandet.ai, and the repository has 411 stars in the two weeks since it opened on 5 October.
The service runs in three shapes. The website is a price browser. The remote MCP server at https://grandet.ai/mcp takes no account and no key, and the readme lists six tools under it: grandet_find_cheapest_api, grandet_find_free_credits, grandet_resolve_model, grandet_quote_evidence, grandet_glossary and grandet_list_apps. The third shape is an installable Agent Skill, which Codex, Claude Code, Cursor and other coding agents can read.
What it reports
Grandet ranks providers by listed price only. The readme says it never calls one site good, never repeats a vendor's own quality claims such as "full strength" or "stable", and never spends credit on behalf of a user. USD credit, CNY credit and cash prices are listed separately, with no cross-currency comparisons.
The lookup covers three cases: the lowest price for a specific model and kind of use (chat, coding, long output), free tiers and trial credit (with a note on whether a card, an ID check or an application is needed), and conditions attached to a price such as a minimum top-up or members-only access. Prices that contradict a site's own text are shown as a reference and the discrepancy is flagged.
How the connect step works
Asked to connect a provider, the Skill walks a user through five steps. The user signs up and tops up on the chosen site in a browser. The user creates a token in the group Grandet names. A terminal command stores the key, with typing hidden so the key never enters the chat. The Skill shows the configuration change it will make, backs up the file and writes only after confirmation. A verification call checks the key, group and model name without spending credit. The readme says "undo" restores the previous setup.
The script that performs the change is scripts/connect.mjs in the repository, also usable as a local MCP server. Other agents get a field card with the names, values and a vendor documentation link, and the user types them in.
The honesty paragraph in the readme
The readme names the limit of a listed price. A listed price is a public number from a vendor's own site. It says nothing about the quality of the answers, the real throughput, or whether a model called "Claude Sonnet 4.6" at a relay site is the same weights as the one at Anthropic. Asked about quality, speed or downgraded models, Grandet reports only what it has on record, and names the case when there is no measurement.
A team buying API capacity across projects pays less in exchange for the time spent reading a price table it did not know existed. The test the readme does not run is whether the cheapest provider for a model is answering with the same model.
Source
Repository: hanzohanzhe/Grandet on GitHub. Website: grandet.ai.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.